I haven't got either of those running myself so I can't hand you working code, but your read on the repos is right and it's worth knowing that's normal rather than something you're doing wrong. Plenty of paper-accompanying repos exist to regenerate the figures and never grow an inference path, and "the weights live somewhere else" is extremely common. Opening an issue asking specifically for a minimal predict-on-one-sequence example is more productive than it sounds -- authors usually do answer that, and you're clearly not the first person to hit it.

Worth considering the alternative though. If what you want is predictions rather than to reproduce a specific paper, training your own on published half-life data may genuinely be faster than getting someone else's code to run. The sequence determinants here aren't exotic: codon usage is one of the strongest single predictors, plus ORF length, 3'UTR length, AU-rich elements, m6A motifs, miRNA seed matches. Saluki's own paper attributes a good chunk of its signal to codon usage and the 3'UTR. Gradient boosting on features like those gets respectable performance against SLAM-seq or TimeLapse-seq derived half-lives, and you'd own the whole pipeline instead of fighting someone else's environment.

Comes down to whether you need state of the art or just something calibrated that you can actually run.